UCLA School of Law

Artificial intelligence and the human mind share a decisionmaking architecture, and antidiscrimination law is built for neither. In both, a producing layer encodes the statistical regularities of an unequal world and shapes the inputs a decision runs on, while an avowing layer deliberates, explains, and sincerely denies discriminatory intent — without being able to observe the processing beneath it. We name this the architecture of disclaim, and we show it is functionally isomorphic across silicon and carbon. Recent computer-science studies of large language models are echoing, sometimes measure for measure, a half-century of implicit bias research on human brains. The parallels are not cosmetic. Models and brains exhibit similar bias effects, with similar resistance to correction. Alignment training and diversity training fail in the same way and for the same reason: each updates what a system says without retraining what it computes. Asked whether it discriminated, either system answers confidently — and the answer is sincere and wrong.The architecture defeats the law's existing toolkit, cluster by cluster. Consciousness (purposeful discrimination) fails because the avowing layer answers honestly. Cause (but-for causation) fails because race travels in a bundle of proxies that no counterfactual can isolate, so the test can convict but never acquit. Consequence (disparate impact) comes closest — it never asks about anyone's interior — but it arrives statutorily confined, constitutionally imperiled, and missing an accepted account of why its disparities matter.A fourth cluster survives. Conduct asks not what the actor intended, observed, or can counterfactually prove, but what it should have done. Two legal traditions are building toward it right now, neither aware of the other. AI governance has converged on assessment, audit, and reasonable response; implicit bias law has replaced Batson's futile hunt for purposeful intent with an epistemically upgraded objective observer. Two construction crews are working with one blueprint. The convergence is not coincidence — it is the signature of the same architecture asserting itself across different substrates — and naming it makes an exchange possible, because each side holds what the other is missing. AI governance holds ex ante duties that attach before any contested decision. Implicit bias law holds the epistemic content without which a reasonableness benchmark is an empty vessel.The exchange composes a new standard: the responsible person — an actor charged with what the science of decisional systems has established, bound by ex ante duties before the decision, and judged on that same knowledge after it. Because the standard interrogates no interior and demands no counterfactual, it reaches the firm whose managers consult a frontier model persistently, informally, and mid-discretion — which is what every firm may be becoming, and what every substrate-specific regime misses. What to do? Judges can begin through evidentiary rulings alone. Legislators can finish, enacting ex ante obligations calibrated to what minds and machines can and cannot do. Two decades ago, one of us asked why we tolerate from our organically grown black boxes what we would never tolerate from synthetic ones. We shouldn’t. The architecture of disclaim means that neither minds nor machines can simply avow their way to fairness — but the law can now hold both to becoming what they claim to be.